# A Tour of Intelligence What an intelligence is, and how its pipeline runs from features and drift to training, evaluation and serving. - Page: https://perceivo.ai/university/intel-01-tour-of-intelligence - JSON: https://perceivo.ai/registry/film/intel-01-tour-of-intelligence.json - Kind: film - Revision: 3 (published 2026-10-18, produced 2026-10-15) - Previous revision 1 (published 2026-10-07): https://perceivo.ai/registry/film/intel-01-tour-of-intelligence/r1.json - Previous revision 2 (published 2026-10-11): https://perceivo.ai/registry/film/intel-01-tour-of-intelligence/r2.json - SDK version: none ## Validation - SDK: not applicable (UI walkthrough; no component or orchestration is offered for download) - Platform: not applicable (UI walkthrough; no component or orchestration is offered for download) ## Transcript [0:00] Introduction to Intelligence, Episode 1: A Tour of Intelligence [0:06] Welcome to Introduction to Intelligence. [0:09] Intelligence is Perceivo's MLOps platform, and it's tightly integrated with the rest of Perceivo. [0:15] You can use an intelligence on its own, or bring it into an orchestration, so that several AI solutions work together toward one goal. [0:24] In this episode, we'll look at what an intelligence is, and how the whole pipeline works: features ingested, drift detected, models trained, evaluated offline, a traffic plan, and predictions served. [0:38] Intelligence lives in the Studio. [0:40] You reach it from the sidebar, right here between orchestrations and components. [0:45] Throughout this series we use one example, a recommender system, and we'll show how Perceivo manages its features, trains its models and serves its predictions. [0:56] We'll walk through it step by step, and then show how automation turns it into an adaptive system. [1:02] Let's open the recommender. [1:04] In Perceivo, an intelligence serves predictions. [1:07] These could be recommendations or classifications, or a time series analysis. [1:14] They could examine x-rays, run an automated factory, or predict when a volcano might erupt. [1:21] If you already have a trained model, you can register it here, load it into an inference service, and start serving predictions right away. [1:28] That's a perfectly good way to use it. [1:31] But an adaptive solution needs more. [1:33] It has to keep taking in new information, make sense of it, and act on it. [1:37] Intelligence automates as much of that as possible, while staying flexible enough for a wide range of machine learning problems. [1:45] When you open an intelligence, you start on this map, which summarizes how your service is configured and how it's running. [1:52] At its heart is a simple flow: features feed training, and training feeds inference. [1:58] Let's take them one at a time. [2:00] Features come first. [2:02] A feature ingest loads new data into Perceivo, and you might run it every day, every hour, or every ten minutes. [2:11] Each run calls a script you provide. [2:13] Perceivo gives you the interface, and your script handles the mechanics of fetching and preparing the data, before storing it in your feature store. [2:22] When new features arrive, you'll want to retrain one or more models. [2:26] It's a bit like a person given new information. [2:29] They have to dedicate some of their brain power to think things over. [2:33] When they have sufficiently refined their understanding, they release the brain power back to the system. [2:39] That is what training does in an intelligence service. [2:43] Perceivo's training module allows you to train multiple models in parallel, harnessing an enormous amount of computing power to contemplate, and then gives it back once it's done. [2:54] Candidate models are scored, and promoted to canary or champion, or deprecated if performance falters. [3:01] The best models are loaded for inference, behind a configurable load balancer. [3:07] You can get a prediction right here, in the Test tab. [3:10] And when you're ready to take it to production, you get your predictions by calling it through the API, or the client SDK. [3:17] Every prediction it serves is recorded, with the model that made it, and a correlation ID that ties it to whatever happens next. [3:35] So far, that's three stages you could run by hand. [3:38] Automation is what connects them. [3:40] Intelligence automation can be driven by a few triggers. [3:45] First, scheduled triggers can be used to run ingestion, training and evaluation on a timetable you choose. [3:52] Perceivo also has chain triggers, which allow you to link the stages based on event completion, so a feature ingest can kick off drift detection, then training, then offline evaluation. [4:04] And with auto-accept turned on, the traffic plan adjustments are applied automatically: promotions loaded, demotions retired. [4:12] Automation is driven by evaluations, which come in three forms, each one answering a different question. [4:19] The first is drift detection, which can run automatically after every ingest. [4:24] It looks for large changes in the population you're serving, and alerts you when it finds one. [4:30] Here you set how much has to change before it counts. [4:33] For example, if your customers have always been mostly women over sixty-five, you may want to use a different model if all your customers are suddenly men under eighteen. [4:43] That's drift detection. [4:45] And in Perceivo, that analysis is built into the platform, and configurable through the UI. [4:52] The second is offline evaluation. [4:54] It kicks off as soon as training finishes. [4:57] The offline evaluation is driven by a script you provide, which scores each model on performance statistics gathered during training. [5:05] Perceivo provides the plumbing, you determine behavior. [5:09] Your evaluation script returns a new traffic plan: a list of models, and the share of traffic each one should get. [5:17] Perceivo implements the plan, automatically promoting and demoting models, launching new ones, retiring the old, and reconfiguring the load balancer to match. [5:28] The third is online evaluation, and this is where the correlation ID comes in. [5:33] How the prediction is generated is only half the story; the other half is what happens when that prediction meets the real world. [5:42] In a recommendation system, once a prediction is given, we'd like to know the outcome. [5:47] Did the customer accept the recommendation, or reject it? [5:51] Did it lead to a click, a purchase, a skip? [5:54] Your online evaluation function is responsible for calling your external system to get those results, and matching them to the predictions stored in Perceivo, by the same correlation ID. [6:05] In this series, a simulator plays the outside world, sending back clicks for the recommendations we serve. [6:12] So the evaluation service compares each new model's click-through rate with the champion's. [6:18] A model that falls well behind is retired, and one that holds up is promoted, producing a new traffic plan. [6:25] Put it all together, and you have the whole pipeline. [6:29] New data comes in and is checked for drift, models retrain, offline evaluation picks the best of them, inference serves them, and online evaluation measures how they really do. [6:40] And then the loop starts again, all without anyone pressing a button. [6:45] Each of these tabs takes one part of that pipeline. [6:48] Over the rest of the series, we'll go through them one at a time, and show how the recommender works. [6:54] So that's Intelligence. [6:55] Features, training and inference, connected by automation, and checked by three kinds of evaluation. [7:02] Together, they turn a model you trained once into a system that keeps learning. [7:07] The Perceivo SDK is open source. [7:10] You'll find it on GitHub, with worked examples, and every component class you've seen today. [7:17] Next time, a quick look at the components behind each of these stages. [7:21] See you there.